Generalized multiple kernel learning with data-dependent priors.

Area of Science:

• Machine Learning
• Probabilistic Modeling
• Data Science

Background:

• Multiple Kernel Learning (MKL) and classifier ensembles are key methods for handling feature sets with varying informativeness or internal inconsistencies.
• Existing MKL methods often struggle with complex data structures and may not fully leverage ensemble learning benefits.

Purpose of the Study:

• To propose a novel probabilistic interpretation of MKL.
• To enhance MKL performance by introducing a data-dependent prior derived from classifier ensembles.
• To develop a unified framework for MKL that accommodates diverse data challenges.

Main Methods:

• A novel probabilistic interpretation of MKL using maximum entropy discrimination with a noninformative prior.
• Introduction of a data-dependent prior based on an ensemble of kernel predictors.
• Development of a hierarchical Bayesian model for simultaneous learning of the prior and classification model.

Main Results:

• The proposed framework establishes a convex optimization problem.
• The model seamlessly incorporates additional information, such as missing views or labels.
• Existing MKL models can be recovered and extended within this framework.
• Demonstrated benefits across supervised, semi-supervised, and partial correspondence tasks.

Conclusions:

• The proposed hierarchical Bayesian MKL framework offers enhanced prediction performance.
• The approach provides a flexible and unified method for various learning problems with multiple views.
• The framework effectively addresses data complexities like missing information and partial correspondence.